Spatial memory at kilometer scale

Walk through a generated city for thirty seconds, turn around, and the street behind you has been redrawn from scratch. Current interactive world models do not remember what they generated.

The root cause is architectural. Most video world models condition on a finite window of recent frames, so anything that leaves the window leaves the world. Attention over longer histories helps, but the cost grows with time rather than with space, which is the wrong axis: revisiting a place after an hour should be cheap if the place hasn't changed.

Our approach keeps an explicit spatial state alongside the generative model: a coarse, queryable map that accumulates what the model has committed to. Generation is conditioned on the map patch around the current viewpoint, and each generated frame writes its commitments back. Memory cost scales with the area explored, not the duration of the session.

At kilometer scale this changes what the models are useful for. Robot navigation policies can be trained on routes with real loop closure, game levels stay consistent between visits, and driving scenarios can return to the intersection where the decision actually mattered.

The open problem is dynamic content. Moving objects break the assumption that the map can treat the world as static. We cover partial fixes, and where they fail, in the full write-up.